Rank Induction for Multiple Instance Learning

Official repository for: Kim et al., "Ranking-Aware Multiple Instance Learning for Histopathology Slide Classification" presented at the Medical Imaging with Deep Learning (MIDL) 2025 conference.

🧠 Overview

Rank Induction is a training strategy for Multiple Instance Learning (MIL) that leverages expert annotations using a ranking constraint—rather than exact attention matching—to guide the model’s focus toward diagnostically meaningful areas. By ranking annotated lesion patches higher than non-lesion patches, our method guides the model to focus on diagnostically meaningful regions without over-constraining attention distribution.

⚠️ Most MIL methods either ignore expert annotations or enforce overly strict attention constraints.
✅ Our method strikes a balance by ranking annotated lesion patches higher than non-lesion ones—offering better interpretability and performance, especially in low-data regimes.

Figure 1. Model performance under data scarcity and interpretation. Left: Performance at varying sampling ratios. Middle: WSI thumbnail. Right: Corresponding attention heatmap.

As shown above, our method maintains stable performance across varying sampling ratios (left) and produces more accurate, interpretable attention maps (right). Red boxes highlight cancerous areas, while blue boxes indicate non-lesion regions.

---

🔬 Key iead

Most classic MIL methods only assume that the presence of at least one positive patch makes a slide positive, with no guidance on how important each patch is. In contrast, Rank Induction uses expert annotations (patches that are known to be lesion or non-lesion) to induce a pairwise ranking preference:

• Lesion patches should have higher attention scores than
• Non-lesion patches.

Formally, for lesion patches ( s_i ) and non-lesion patches ( s_j ), we want:

$$s_i > s_j \quad \text{for all} \; (i, j) \;\text{where} \; y_i = 1, \; y_j = 0.$$

Pairwise Probability

To implement this, we convert the score difference into a pairwise probability ( P_{i,j} ):

$$P_{i,j} \;=\; \frac{1}{1 + \exp\big[-\sigma ( s_i - s_j - m )\big]},$$

where

• $s_i$ and $s_j$ are the raw attention score (before softmax),
• $\sigma$ is a scaling factor,
• $m$ is a margin to encourage a significant gap between lesion and non-lesion attention scores.

Rank Loss

We define the Rank Loss $\mathcal{L}\mathrm{rank}$ by comparing the predicted pairwise probability ($P{i,j}$) against the ground-truth preference ($\bar{P}_{i,j} \in {0,1} $) (which indicates which patch should be ranked higher):

$$\mathcal{L}_{\text{rank}} = \frac{1}{|\mathcal{P}|} \sum_{(i,j)\in \mathcal{P}} -\bar{P}_{i,j} \,\log P_{i,j} \;-\; \bigl(1 - \bar{P}_{i,j}\bigr)\,\log\bigl(1 - P_{i,j}\bigr),$$

where ($\mathcal{P}$) contains all valid (lesion–non-lesion) patch index pairs.
Finally, the slide-level classification objective (e.g., binary cross-entropy) and the rank loss are jointly optimized.

---

📊 Results

Camelyon16

Method	AUROC	AUPRC
AB-MIL	0.740 ± 0.146	0.730 ± 0.183
Attention Induction	0.743 ± 0.142	0.727 ± 0.179
Rank Induction	0.836 ± 0.044	0.851 ± 0.036

DigestPath

Method	AUROC	AUPRC
AB-MIL	0.992 ± 0.002	0.989 ± 0.004
Attention Induction	0.993 ± 0.002	0.990 ± 0.003
Rank Induction	0.995 ± 0.001	0.993 ± 0.002

🧪 Reproducibility
To reproduce the above results, please refer to the Jupyter notebook provided in:

result/
├── 0_loss.ipynb
├── 1_performance.ipynb
├── 2_posthoc.ipynb
├── 3_data_scarcity.ipynb
├── 4_granularity.ipynb
└── 5_digestpath.ipynb

👨‍🔬 Authors

Name	ORCID	Email	Affiliation	Notes
Ho Heon Kim	0000-0001-7260-7504	hoheon0509@mf.seegene.com	$^{1}$ AI Research Center, Seegene Medical Foundation	Contributed equally
Gisu Hwang	0000-0003-1046-9286	gshwang@mf.seegene.com	$^{1}$ AI Research Center, Seegene Medical Foundation	Contributed equally
Won Chang Jeong	0009-0008-1931-5957	jeongwonchan53@mf.seegene.com	$^{1}$ AI Research Center, Seegene Medical Foundation
YoungSin Ko	0000-0003-1319-4847	noteasy@mf.seegene.com	$^{1,2}$ AI Research Center / Pathology Center, Seegene Medical Foundation	Corresponding author

📍 Affiliations

• $^{1}$ AI Research Center, Seegene Medical Foundation, 320 Cheonho-daero, Seoul, South Korea
• $^{2}$ Pathology Center, Seegene Medical Foundation, 320 Cheonho-daero, Seoul, South Korea